Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.

4, August 2014
DOI : 10.5121/sipij.2014.5402 15

SYMBOLIC REPRESENTATION AND
RECOGNITION OF GAIT : AN APPROACH
BASED ON LBP OF SPLIT GAIT ENERGY
IMAGES
Mohan Kumar H P
1
and Nagendraswamy H S
2
1
Department of MCA, PES College of Engineering, Mandya, India
2
Department of studies in Computer Science, Manasagangothri, Mysore, India


ABSTRACT

Gait is one of the biometric techniques used to identify an individual from a distance by his/her walking
style. Gait can be recognized by studying the static and dynamic part variations of individual body contour
during walk. In this paper, an interval value based representation and recognition of gait using local
binary pattern (LBP) of split gait energy images is proposed. The gait energy image (GEI) of a subject is
split into four equal regions. LBP technique is applied to each region to extract features and the extracted
features are well organized. The proposed representation technique is capable of capturing variations in
gait due to change in cloth, carrying a bag and different instances of normal walking conditions more
effectively. Experiments are conducted on the standard and considerably large database (CASIA database
B) and newly created University of Mysore (UOM) gait dataset to study the efficacy of the proposed gait
recognition system. The proposed system being robust to handle variations has shown significant
improvement in recognition rate.

KEYWORDS

Gait Recognition, Gait energy image, Interval-valued features, Local binary pattern, Representation,
Symbolic data, Similarity measure
1. INTRODUCTION
Gait is a manner of walking of an individual. Gait recognition is a method of identifying or
verifying an individual based on his/her walking style. Gait recognition has gained much attention
in recent years in the field of computer vision because of its ability to deal with low resolution
images captured at a considerable distance away from the camera without the individuals
cooperation. Gait can be used in situations when other biometric traits such as face, iris and
fingerprint information do not have sufficient resolution for recognition. Gait as a biometric
source can be used in some monitoring applications for early warnings if any, suspicious threats
are found. Though several techniques have been proposed for efficient gait recognition, the
methods have some limitations in terms of time complexity, storage requirements, recognition
rate and works under some constraints. So there is a scope for exploring further the robust
techniques for gait recognition.


Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
16
2. RELATED WORK

Several attempts have been made by many researchers to provide an efficient and effective gait
recognition system. Following are the few interesting works found in the literature in this
direction.

Han et al. [1] introduced Gait energy image, which consist of spatio-temporal information.
Recognition is achieved by combining real and synthetic templates (poses). USF Gait database is
used for experimentation and the study shows satisfactory results. Pratik chattopadhyay et al. [2]
proposed pose depth volume (PDV) gait feature for frontal view gait recognition. The method
combines both colour and depth information in the extracted feature. A PDV is constructed by
averaging voxel volumes of depth key pose frames. It consists of both shape and depth variations
over each depth key pose in a gait cycle of individual walking sequence. Dataset used is captured
in an indoor environment using Microsoft kinect. Experiment is conducted on 30 subjects and
author claimed acceptable results. Jeevan et al. [3] proposed gait recognition using Pal and Pal
Entropy. PCA is applied on the feature matrix and Support Vector Machine is used for
classification. The author compares the result with Shannon Entropy and achieved reasonable
result. Sudeep Sarkar et al. [4] proposed a baseline algorithm which consists of 12 experiments.
Recognition is achieved through frame correlations between successive frames using Tanimoto
similarity measure. Recognition rate for 12 experiments varies from 78% on the easiest to 3% on
the hardest experiment. AmitKale et al. [5] proposed a method for gait recognition, which uses
the binary silhouette width as features. The feature incorporates both structural and dynamic
information of an individual. Experimental results have shown that the method has better
recognition rate for side view compared to frontal view. Dupuis et al. [6] proposed the Random
Forest (RF) algorithm which is based on the bootstrap-aggregating concept to rank features
importance to address the problem of high dimensional feature space in model-free approach. In
order to efficiently search throughout subspaces, they have applied a backward feature
elimination search strategy. Authors have claimed that their approach can greatly reduce the
complexity of the classification problem while achieving fair correct classification rates when gait
is captured with unknown conditions.

Most of the techniques proposed in the literature for gait recognition incorporate the features
extracted from each silhouette of a gait sequence. Since each gait sequence has considerably more
number of silhouettes, the process of feature extraction, representation and matching for gait
recognition becomes complex and time consuming for real time applications. In order to
overcome this limitation, in this work, the idea of GEI has been explored and the suitability of
texture feature extraction technique such as LBP is studied. GEI image is divided into four equal
parts to localize the variations in each part and LBP features are extracted from each part. The
extracted features are organized in a clock wise direction to preserve the sequence. Sequence
preservation is very much essential for matching and recognition.

The gait of an individual person may vary due to change in cloth, change in shoe, change in
surface, carrying a bag etc. Also it has been observed that there will be some variations even in
normal walking condition at different point of time. The conventional data analysis techniques
may fail to capture such variations effectively. From the literature survey, we understand that the
concept of symbolic data analysis has been well studied in the field of cluster analysis [7, 8, 9, 10,
11, 12], shape analysis [13] and signature biometric applications [14]. Also suitability of
symbolic data analysis approach for gait recognition is attempted recently in [15, 16, 17, 18].
These unconventional techniques have proved that they outperform the conventional techniques
in terms of performance and uncertainty. Thus, we propose to incorporate the concept of
symbolic data analysis particularly the interval type data to capture the variations and to
effectively represent the gait information in the knowledge base used for the purpose of
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
17
recognition. Performance of the proposed gait recognition system is studied by conducting
experiments on the CASIA B and UOM gait databases. Rest of the paper is organized as follows.
Section 3 presents proposed methodology for gait recognition. Experimental results are presented
in section 4, followed by discussion and conclusion in section 5.

3. PROPOSED METHODOLOGY

The proposed method of gait recognition system uses GEI and LBP technique to extract features
for gait representation. The method involves the following steps: In the first step, Gait Energy
Image (GEI) is generated using sequence of silhouettes of a gait cycle. In the second step, GEI is
divided into four equal parts. In the third step, LBP operator is applied to extract features from
each part of GEI separately and then concatenated for representation. The feature extraction
process is applied to all instances (different covariates such as wearing coat, carrying a bag and
different normal conditions) of a subject and the corresponding features are consolidated to form
an interval-valued feature vector representing a subject in the gait knowledgebase. In the fourth
step, a suitable similarity measure is used to match probe gait with reference gait and a matching
score is computed for recognition. The following subsections describe these steps in detail.

3.1. Gait energy image
In order to reduce the complexity in characterizing a gait, the silhouettes of a gait cycle are
aggregated using the idea proposed by [1] to produce a single image called Gait Energy Image
(GEI). GEI for a gait cycle is computed by taking the average of all silhouettes over a gait cycle
and is defined by

( ) ( ) ) 1 ( , ,
1
,
1

=
= =
T
t
t y x I
T
y x G GEI

Where T is the number of frames in a gait cycle, I is a gait silhouette image, x and y are pixel
coordinates and t is the frame number in a sequence of silhouettes of a gait cycle.

Fig. 1 shows the examples of GEIs computed for same subject in three different instances (a)
carrying a bag (b) with coat and (c) normal and Fig. 2 shows the same GEIs divided into four
equal parts.

(a) (b) (c) (a) (b) (c)
Figure 1 Figure 2

3.2. Feature extraction
Features are extracted from GEI to characterize the gait. The idea of LBP proposed in [19] is used
to extract the features. LBP is a gray scale texture operator which describes local texture pattern
with a binary code. A binary pattern number for a central pixel is obtained by comparing the gray
value of central pixel with gray values of its neighbourhood pixels as shown below.

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
18

) 2 ( 2 ) ( ) , (
1
0
,
p
P
P
c p c c R P
g g s y x LBP

=
=



) 3 (
0 , 0
0 , 1
) (

<

=
x
x
x s

Where
c
g is the gray value of the center pixel ( )
c c
y x , and
p
g is the gray value of its
neighbours in the P sampling points plane. P is the number of neighbours and R is the radius of
the neighbourhood.

In this work, we used uniform LBP pattern with P = 8 and R = 1, to extract 59 bin values as
features. From each part of GEI, 59 feature values are extracted. Since, the feature values are
extracted from four different parts of GEI, totally 236 feature values are used to represent each
gait of a subject. The more theoretical description about uniform LBP pattern is found in [20].
3.3. Representation
Since the gait of a person varies slightly due to change in carrying conditions, change in clothes
and different normal conditions, the LBP features obtained for these instances (covariates) also
contain variations. These variations are handled by consolidating the features in the form of an
interval type data as explained below.
Let [ ] ) 4 ( , , ,
2 1 N I
S S S S S =
be the N number of subjects.
Let { } ) 5 ( , , , , ,
2 1 n
I
j
I I I I
s s s s S =

be the n instances of the subject
I
S .

LBP features extracted separately from four different parts of GEI (Fig. 2) in a clockwise
direction, numbered as 1,2,3 and 4 for an instance
j
I
s of subject
I
S is given as


{ } { } { } { } [ ] ) 6 ( 4 , 3 , 2 , 1 F F F F s
j
I
=
Where

{ } ) 7 ( 1 , , , 1 , , 1 , 1 1
Im 2 1
j j
Ik
j
I
j
I
f f f f F =

{ } ) 8 ( 2 , , , 2 , , 2 , 2 2
Im 2 1
j j
Ik
j
I
j
I
f f f f F =

{ } ) 9 ( 3 , , , 3 , , 3 , 3 3
Im 2 1
j j
Ik
j
I
j
I
f f f f F =


{ } ) 10 ( 4 , , , 4 , , 4 , 4 4
Im 2 1
j j
Ik
j
I
j
I
f f f f F =


and m is the number of LBP features.
The mean and standard deviation of
th
k feature of first part F1 of all the instances of
I
S is given
by
( ) ) 11 ( 1 , 1 , , 1 , 1 1
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f mean =


( ) ) 12 ( 1 , 1 , , 1 , 1 1
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f std =


Similarly the mean and standard deviation of
th
k feature of second (F2), third (F3) and fourth
(F4) part of all the instances of
I
S is computed as follows:
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
19

( ) ) 13 ( 2 , 2 , , 2 , 2 2
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f mean =


( ) ) 14 ( 2 , 2 , , 2 , 2 2
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f std =


( ) ) 15 ( 3 , 3 , , 3 , 3 3
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f mean =


( ) ) 16 ( 3 , 3 , , 3 , 3 3
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f std =


( ) ) 17 ( 4 , 4 , , 4 , 4 4
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f mean =


( ) ) 18 ( 4 , 4 , , 4 , 4 4
2 1 n
Ik
j
Ik Ik Ik Ik
f f f f std =

The minimum and maximum value of the
th
k feature of first part F1 of all the instances of
I
S is
given by

) 19 ( 1 1 1
Ik Ik Ik
f =

) 20 ( 1 1 1
Ik Ik Ik
f + =
+


Similarly the minimum and maximum value of
th
k feature of second (F2), third (F3) and fourth
(F4) part of all the instances of
I
S is given by

) 21 ( 2 2 2
Ik Ik Ik
f =

) 22 ( 2 2 2
Ik Ik Ik
f + =
+


) 23 ( 3 3 3
Ik Ik Ik
f =

) 24 ( 3 3 3
Ik Ik Ik
f + =
+


) 25 ( 4 4 4
Ik Ik Ik
f =

) 26 ( 4 4 4
Ik Ik Ik
f + =
+


Thus, the reference gait of a subject
I
S (I = 1,2,..,N) in the knowledge base is represented in the
form of interval-valued type symbolic feature vector as follows:


{ } { } { } { } [ ] ) 27 ( 4 , 3 , 2 , 1
I I I I I
RF RF RF RF S =


Where

{ } ) 28 ( 1 , , 1 , , 1 , 1 1
Im 2 1
rf rf rf rf RF
Ik I I I
=


{ } ) 29 ( 2 , , 2 , , 2 , 2 2
Im 2 1
rf rf rf rf RF
Ik I I I
=


{ } ) 30 ( 3 , , 3 , , 3 , 3 3
Im 2 1
rf rf rf rf RF
Ik I I I
=


{ } ) 31 ( 4 , , 4 , , 4 , 4 4
Im 2 1
rf rf rf rf RF
Ik I I I
=
and

{ } ) 32 ( 1 , 1 1
+
=
Ik Ik Ik
f f rf


{ } ) 33 ( 2 , 2 2
+
=
Ik Ik Ik
f f rf


{ } ) 34 ( 3 , 3 3
+
=
Ik Ik Ik
f f rf


{ } ) 35 ( 4 , 4 4
+
=
Ik Ik Ik
f f rf

The probe gait of a subject
I
S is also characterized with LBP features as described earlier but
with only one instance (covariate). Thus the feature vector representing probe gait is a crisp
vector as follows:

{ } { } { } { } [ ] ) 36 ( 4 , 3 , 2 , 1 PF PF PF PF s
P
=

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
20
Where

{ } ) 37 ( 1 , , 1 , , 1 , 1 1
2 1 m k
f f f f PF =


{ } ) 38 ( 2 , , 2 , , 2 , 2 2
2 1 m k
f f f f PF =


{ } ) 39 ( 3 , , 3 , , 3 , 3 3
2 1 m k
f f f f PF =


{ } ) 40 ( 4 , , 4 , , 4 , 4 4
2 1 m k
f f f f PF =

Table 1 shows only the 1
st
, 2
nd
and 59
th
feature values extracted from part-1, part-2, part-3 and
part-4 of GEI of different instances (covariates) of a subject
I
S as an example. Table 2 shows the
interval-valued representation of the 1
st
, 2
nd
and 59
th
feature values of all the four parts of GEI for
a subject
I
S .


















3.4. Similarity computation
In order to recognize a probe gait
P
S , features are extracted from the probe gait as discussed in
sub section 3.2 of section 3 and represented as shown in Equation 36. As an example, only a few
Table 1. Crisp feature values extracted from first, second, third and fourth GEI
part of different instances (covariates) of a subject.
GEI Part 1 GEI Part 2
Feature No 1 2 59 1 2 59
Bag1 24 36 161 24 48 129
Cloth1 19 34 159 28 52 130
Normal1 11 30 152 21 41 119
Normal2 13 29 157 17 39 123
Normal3 10 26 140 19 44 120
GEI Part 3 GEI Part 4
Feature No 1 2 59 1 2 59
Bag1 18 57 209 27 92 112
Cloth1 18 63 196 29 86 116
Normal1 13 52 189 20 77 108
Normal2 11 49 200 26 83 103
Normal3 16 54 192 24 84 99

Table 2. Interval valued features representing a subject.
GEI Part 1 GEI Part 2
Feature No Interval Feature No Interval
1 [9.45 , 21.34] 1 [17.47, 26.12]
2 [27.00, 35.00] 2 [39.53, 50.06]

59 [145.39, 162.20] 59 [119.13, 129.26]
GEI Part 3 GEI Part 4
Feature No Interval Feature No Interval
1 [12.08, 18.31] 1 [21.77, 28.62]
2 [49.66, 60.33] 2 [78.98, 89.81]

59 [189.40, 204.99] 59 [100.79, 114.40]

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
21
feature values of type crisp of a probe gait
P
S is shown in Table 3. The obtained crisp feature
vector of a probe gait is compared with the symbolic feature vectors of reference gaits Equation
27 in the gait knowledge-base and a matching score is computed for recognition.








The Similarity measure suggested in [21] is found to be suitable and hence used for computing
similarity between reference gaits and probe gait as described below.


( ) [ ] ( )
N to I for
x f x f x f sim S S arity TotalSimil
Ik Ik k
m
k x
I p
1
) 41 ( , , ,
1
4
1
=
=
+
= =



Where


[ ] ( )
( )
) 42 (
1
1
,
1
1
max
1
, ,
'

|
|
|

\
|
+ +

=
+
+
+
otherwise
x f x f x f x f
x f x f and x f x f if
x f x f x f sim
Ik k Ik k
Ik k Ik k
Ik Ik k

When
k
x f lies between the interval, the similarity value will be 1. Otherwise, the similarity
value depends on the extent to which the
k
x f value is closer to either lower limit

Ik
x f or the
upper limit
+
Ik
x f . The similarity between the probe gait and reference gait of all the subjects
I
S (I = 1,2,..,N) in the knowledge-base is computed and is used at the time of identification as
discussed in section 4.
4. EXPERIMENTS

In order to study the performance of the proposed method of gait recognition, two experiments
were carried out in this work. In the first experiment, training set is composed of a mixture of gait
sequences under different covariate conditions. In the second experiment, the training set contains
gait sequences of subjects walking under similar covariate conditions. The gait silhouettes used
are in 90 degree (side view) viewing angle as this view provides more gait information than the
silhouettes taken from other view angles.

4.1. Datasets used in the experiments
We have conducted experiments on the standard CASIA B dataset [22] and on our newly created
University of Mysore (UOM) gait dataset. The CASIA dataset consists of 124 individuals
(subjects) with three covariates such as view angle, carrying condition and wearing coat. Each
subject consists of 10 series, out of which 2 series are walking sequences carrying a bag, 2 series
are walking sequences wearing different clothes and 6 series are in normal conditions.

Due to unavailability of other standard data sets (except CASIA) for us, we created UOM gait
dataset to check the efficacy of the proposed system with more covariates. The UOM dataset
Table 3. Few crisp feature values representing probe gait.
Feature No GEI Part 1 GEI Part 2 GEI Part 3 GEI Part 4
1 20 29 14 29
2 37 43 59 77

59 157 116 201 110
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
22
consists of 66 subjects of 8 sequences per subject captured in 90
0
(side) view with six covariates
such as surface, foot wear, carrying bag, holding an object, cloth change and elapsed time.
Among 66 subjects, 41 are male and 14 are female subjects. A male subject walking holding an
object like water bottle or a female subject walking holding a purse is a new covariate introduced
newly by us in the UOM gait dataset and is realistic in real world scenario. Each subject includes
all the six covariates and consists of 8 series, out of which 4 series are normal walking, 2 series
carrying a bag and 2 series holding an object. The UOM dataset contains a total of 528 sequences
of 66 subjects.

We have measured the performance of the proposed gait recognition system using cumulative
match scores (CMS) suggested in [23]. The task of recognition is to identify a given probe gait to
be one of the reference gait.
4.2. Experimental results on CASIA B database
4.2.1. Experiment I
In this experiment, first series of carrying a bag named as B1 (bag1), first series of coat named as
C1 (cloth1) and first three different normal walking series named as N1 (normal1), N2 (normal2)
and N3 (normal3) are used for training. Second series of carrying a bag named as B2 (bag2),
second series of coat named as C2 (cloth2) and rest of the series of normal walking are named as
N4 (normal4), N5 (normal5) and N6 (normal6) are used for testing. Table 4 shows the
identification rate of the proposed methodology at rank 1, 5 and 10. The Cumulative Match curve
for the proposed system in Fig. 3 shows that the performance at rank 1 is the correct classification
rate (CCR) and we have achieved average CCR of 88.99%.







4.2.2. Experiment II
In the second experiment, only first four different normal walking series named as N1 (normal1),
N2 (normal2), N3 (normal3) and N4 (normal4) are used for training. Other two series of carrying
a bag named as B1 (bag1), B2 (bag2), two series of coat named as C1 (cloth1) and C2 (cloth2)
and last two different normal walking series named as N5 (normal5), and N6 (normal6) are used
for testing. Table 5 shows the identification rate of the proposed methodology at rank 1, 5 and 10.
The Cumulative Match curve for the proposed system in Fig. 4 shows that the performance at
rank 1 is the correct classification rate (CCR) and we have achieved average CCR of 79.01%.


Table 4. Identification rates at different ranks in the proposed
approach.
Probe Identification rate/Rank (%)
1 5 10
N4 93.33 100.00 100.00
N5 94.16 100.00 100.00
N6 91.66 100.00 100.00
C2 83.33 92.50 98.33
B2 82.50 90.80 97.50
Average CCR 88.99

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
23











Figure 3. Cumulative Match score








Table 6 shows the recognition performance (% of average correct classification rate at rank 1) of
experiment I and experiment II as discussed in sub section 4.2.1 and 4.2.2 respectively of the
proposed method and other methods reported in [6] and [3].









Table 5. Identification rates at different ranks in the
proposed approach.
Probe Identification rate/Rank (%)
1 5 10
N5 96.66 100.00 100.00
N6 95.83 100.00 100.00
C1 65.80 71.60 79.16
C2 68.30 74.16 78.30
B1 73.33 79.16 85.80
B2 74.16 80.00 88.30
Average CCR 79.01

0 1 2 3 4 5 6 7 8 9 10
0
10
20
30
40
50
60
70
80
90
100
Cumulative match score
Rank
I
d
e
n
t
i
f
i
c
a
t
i
o
n

R
a
t
e


bag2
cloth2
normal4
normal5
normal6

Table 6. Average CCR of other method and proposed method for side view
(90
0
)
Approaches Average CCR (%) Database
Y. Dupuis [6] 78.8
CASIA B
Jeevan [3] 70.24
Proposed method exp I 88.99
Proposed method exp II 79.01
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
24

Figure 4. Cumulative Match score
4.3. Experimental results on UOM database
4.3.1. Experiment I
In the first experiment, first series of carrying a bag named as B1 (bag1), first series of holding an
object named as O1 (object1), first series of normal walking named as N1 (normal1), and third
series of normal walking named as N3 (normal3) are used for training. Second series of carrying a
bag named as B2 (bag2), second series holding an object named as O2 (object2), second series of
normal walking named as N2 (normal2), and fourth series of normal walking named as N4
(normal4) are used for testing. Table 7 shows the identification rate of the proposed methodology
at rank 1, 5 and 10. The Cumulative Match curve for the proposed system in Fig. 5 shows that the
performance at rank 1 is the correct classification rate (CCR) and we have achieved average CCR
of 92.79%.










Table 7. Identification rates at different ranks in the proposed
approach.
Probe Identification rate/Rank (%)
1 5 10
N2 95.45 98.48 100.00
N4 98.48 100 100.00
O2 90.90 93.93 98.48
B2 86.36 90.90 95.45
Average CCR 92.79

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
25

Figure 5. Cumulative Match score
4.3.2. Experiment II
In this experiment, first series of normal walking named as N1 (normal1) and third series of
normal walking named as N3 (normal3) are used for training. First series of carrying a bag named
as B1 (bag1), second series of carrying a bag named as B2 (bag2), first series of holding an object
as O1 (object1), second series holding an object named as O2 (object2), second series of normal
walking named as N2 (normal2), and fourth series of normal walking named as N4 (normal4) are
used for testing. Table 8 shows the identification rate of the proposed methodology at rank 1, 5
and 10. The Cumulative Match curve for the proposed system in Fig. 6 shows that the
performance at rank 1 is the correct classification rate (CCR) and we have achieved average CCR
of 86.35%.











Table 8. Identification rates at different ranks in the
proposed approach.
Probe Identification rate/Rank (%)
1 5 10
N2 98.48 100.00 100.00
N4 96.96 100.00 100.00
O1 81.81 84.84 93.93
O2 84.84 87.87 95.45
B1 78.78 83.33 92.42
B2 77.27 80.30 89.39
Average CCR 86.35
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
26

Figure 6. Cumulative Match score

5. DISCUSSION AND CONCLUSION

We have conducted two experiments on both CASIA B and UOM data sets. In the first
experiment on CASIA B data set, the proposed gait recognition system has achieved 88.99% of
average correct classification rate (ACCR) and in the second experiment, 79.01% of ACCR is
reported. A significant difference in the percentage of ACCR is observed in these experiments. In
the first experiment, all the covariates of a subject such as change in cloth, carrying a bag and
different instances of normal walking conditions are considered during training but in the second
experiment, only different instances of normal walking conditions are considered during training.
In both the cases, the variations are captured in the form of interval valued type symbolic data.
Since, all the covariates are considered in the first experiment, the ACCR (88.99%) is high when
compared to the ACCR (79.01%) obtained in the second experiment. This clearly shows that it is
very much essential to consider all possible covariates of a gait during training and capture the
variations effectively to improve the recognition rate. Though, we have considered more
covariates for a subject during training, there is only one reference gait information for a
particular subject is stored in the knowledge-base and hence the total number of reference gaits is
same in both the experiments for CASIA B data set. This is possible because of symbolic
representation as discussed earlier. The above argument is also true for UOM data set. Thus, the
proposed representation technique is capable of capturing variations without increasing the size of
the knowledge-base. Also observe that the ACCR (88.99% for different covariates in training set)
and ACCR (79.01% for similar covariates in training set) obtained from the proposed system for
CASIA B data set is significantly high when compared to the ACCR (78.80%) reported by
Dupuis [6] and ACCR (70.24%) reported by Jeevan [3] for similar covariates in training set for
the same data set.
The proposed approach to gait recognition is found to be robust in capturing variations of gait due
to different covariates. The concept of GEI has drastically reduced the complexity of algorithm in
terms of space and time needed for gait recognition. The idea of extracting LBP features from
split GEIs has effectively captures the local variations of gait information. Experiments conducted
on two databases of considerably large size show that the proposed method is robust in capturing
variations and the results obtained are encouraging and are comparable with one of the
contemporary method. Some other challenging issues such as arbitrary viewing angle,
unconstrained dress code and environment will be addressed in our future work.
Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
27
REFERENCES
[1] Ju Han and Bir Bhanu, Individual Recognition using Gait Energy Image, IEEE Transactions on
Pattern analysis and Machine Intelligence, Vol 28, NO 2, February 2006.
[2] Pratik Chattopadhyay, Aditi Roy, Shamik Sural and Jayanta Mukhopadhyay, Pose Depth Volume
extraction from RGB-D streams for frontal gait recognition, J.Vis. Commun. Image R.(2013),
http://dx.doi.org/10.1016/j.jvcir.2013.02.010.
[3] Jeevan M, Neha Jain, M.Hanmandlu and Girija Chetty, Gait Recognition based on Gait Pal and
Entropy Image, 978-1-4799-2341-0/13 IEEE.
[4] Sudeep Sarkar, P. Jonathon Philips, Zongyi Liu, Isidro Robledo Vega, Patrick Grother and Kevin W.
Bowyer, The HumanID Gait Challenge Problem: Data sets, Performance, and Analysis, IEEE
Trans. Pattern Analysis and Machine Intelligence, Vol. 27, No.2, February (2005).
[5] Amit Kale, Aravind Sundaresan, A.N.Rajagopaln, Naresh P.Cuntoor,Amit K. Roy-chowdhury,
Volker Kruger and Rama Chellappa, Identification of Humans Using Gait, IEEE Trans on Image
Processing, Vol.13, NO.9, September (2004).
[6] Y. Dupuis, X.Savatier and P.Vasseur, Feature subset selection applied to model-free gait
recognition, Image and vision computing 31 (2013) 580-591.
[7] T.Denoeux and M.Masson, Mutidimensional scaling of interval valued dissimilarity data, Pattern
Recognition Letters 21, 83-92 (2000).
[8] K.C.Gowda and E.Diday, Symbolic clustering using a new dissimilarity measure, Pattern
Recognition 24(6), 567-578 (1991).
[9] K.C.Gowda and E.Diday, Symbolic clustering using a new similarity measure, IEEE Trans. SMC
22(2), 368-378 (1992).
[10] K.C.Gowda and T.V.Ravi, Divisive clustering of symbolic objects using the concepts of both
similarity and dissimilarity, Pattern Recognition Letters 28(8), 1277-1282 (1995b).
[11] D.S Guru, and B.B.Kiranagi, Multivalued type dissimilarity measure and concept of mutual
dissimilarity value useful for Clustering symbolic patterns, Pattern Recognition 38(1), 151-156
(2005).
[12] D.S Guru, B.B.Kiranagi and P.Nagabhushan, Multivalued type proximity measure and concept of
mutual similarity value useful for clustering symbolic patterns, Pattern Recognition Letters 25(10),
1203-1213 (2004).
[13] D.S Guru and Nagendraswamy H S, Symbolic representation of two-dimensional shapes, Pattern
Recognition Letters 28 144-155 (2007).
[14] D.S Guru and H N Prakash, Online Signature Verification and Recognition:An Approach Based on
Symbolic Representation, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.
31, NO 6, June (2009).
[15] Mohan Kumar H P and Nagendraswamy H S, Gait Recognition: An Approach Based on Interval
valued Features, 978-1-4673-2906-4/13 IEEE.
[16] Mohan Kumar H P and Nagendraswamy H S, Change Energy Image for Gait Recognition: An
Approach based on Symbolic Representation, International Journal of Image, Graphics and Signal
Processing (IJIGSP), Vol. 6, NO. 4, pp. 1-8, 2014.
[17] Mohan Kumar H P and Nagendraswamy H S, Gait Recognition Based on Symbolic Representation,
International Journal of Machine Intelligence, Vol. 3, Issue 4, December 2011, pp-295-301 (2011).
[18] Mohan Kumar H P and Nagendraswamy H S, LBP for Gait Recognition: A Symbolic Approach
based on GEI plus RBL of GEI, International Conference on Electronics and Communication
System (ICECS-2014), Coimbatore, India, Feb 13-14 (2014).
[19] Timo Ojala, Matti Pietikainen and Topi Maenpaa, Multiresolution Gray-Scale and Rotation Invariant
Texture Classification with Local Binary Patterns, IEEE Trans on Pattern and Machine Intelligence,
VOL 24, NO 7, July 2002.
[20] Zhenhua Guo, Lei Zhang and David Zhang, Rotation Invariant Texture Classification using LBP
Variance (LBPV) with Global Matching, Pattern Recognition 43 (2010) 706-719.
[21] H.S.Nagendraswamy and D.S.Guru, A new method of representing and matching two dimensional
shapes, International Journal of Image and Graphics, Vol 7, NO.2, April (2007).
[22] CASIA dataset , Center for Biometrics and Security Research, http://www.cbsr.ia.ac.cn.
[23] P.J.Philips, H. Moon, S Rizvi, and P.Rauss, The FERET Evaluation Methodology for Face-
Recognition Algorithms, IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 22 No 10, pp.
1090-1104, Oct 2000.

Signal & Image Processing : An International Journal (SIPIJ) Vol.5, No.4, August 2014
28
AUTHORS

Mohan Kumar H P obtained MCA and M.Sc Tech by research from University of Mysore,
India in 1998 and 2009 respectively. He is working as Assistant professor in department of
MCA, PES College of Engineering, Mandya, Karnataka, India. He is currently working
towards his Ph.D degree in the area of computer vision at University of Mysore.

Nagendraswamy H S obtained his M.Sc and Ph.D degrees from University of Mysore,
India in 1994 and 2007 respectively. He is currently working as Associate Professor in the
Department of Studies in Computer Science, University of Mysore, Manasagangothri,
Mysore, Karnataka, India. His focused areas of research include Shape analysis, Texture
analysis, Precision agriculture, Symbolic data analysis, Fuzzy theory, Biometric, and Video
analysis.